A semantic parsing model training method and system

CN117591545BActive Publication Date: 2026-10-16GUILIN UNIV OF ELECTRONIC TECH
2 Cites 0 Cited by

Patent Information

Application Number
CN202311600654.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-11-28
Publication Date
2026-10-16
Estimated Expiration
2043-11-28

AI Technical Summary

Technical Problem

[0004]为了克服现有的数据表预训练方法忽视了NL问题与数据库模式之间的交互，使得训练模型输出的结果准确性较低的问题，本发明提供了一种语义解析模型训练方法及系统

Benefits of technology

[0021] The beneficial effects of this invention are as follows: Sample vectors corresponding to words with pre-masked question tags are used to train a pre-defined retention network. During training, each sample vector undergoes feature extraction sequentially through multiple retention layers, and the predicted masking tags corresponding to the masked question tags in each sample vector are output. For each predicted masking tag, pattern dependency prediction is performed from the table patterns corresponding to all sample vectors, determining each predicted directed edge and its corresponding predicted label for each predicted masking tag. Based on the second loss value of each predicted directed edge and the third loss value of the corresponding predicted label, the predicted directed edge with the strongest pattern dependency for each predicted masking tag in the table pattern and its corresponding predicted label can be determined, thereby accurately finding the pre-defined word corresponding to each predicted masking tag in the table pattern. Furthermore, the first loss value, each second loss value, and the corresponding third loss value for each predicted masking tag can be used to determine whether the trained retention network meets the accuracy requirements. If it does not meet the requirements, the parameters in the retention network are adjusted until the trained retention network meets the accuracy requirements. Thus, the interaction between the predicted masking tags and the database table patterns is considered during model training, improving the accuracy of the output results of the trained model.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader

Abstract

The present application relates to a kind of semantic analysis model training method and system, method includes: obtaining to-be-trained vector sequence;The training vector sequence is input into the first reserved layer, and the shield mark corresponding to the prediction shield mark of each word is output by the last reserved layer;Determine the prediction directed edge and prediction label between each column data in pre-set table mode and each prediction shield mark;Determine each fifth loss value corresponding to each prediction shield mark, if each fifth loss value is less than preset value, then each fifth loss value less than preset value corresponding reserved network is used as target semantic analysis model, if there is fifth loss value greater than or equal to preset value in each fifth loss value, then the parameter of reserved network is adjusted, until each fifth loss value is less than preset value.The existing data table pre-training method ignores the interaction between NL problem and database mode, so that the accuracy of the result output by the training model is low.
Need to check novelty before this filing date? Find Prior Art